claude-code - 💡(How to fix) Fix Bootstrapped Agentic OS as Claude Code Default Onboarding [1 participants]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

Proposal: Bootstrapped Agentic OS as Claude Code Default Onboarding Submitted by Peter Parise — GitHub: Raldan951

The Problem Claude Code ships with one onboarding primitive: CLAUDE.md. It's the right first step, but it's where the guidance ends. The gap between "I installed Claude Code" and "I have a working memory system with session continuity" is enormous — and most users never close it. They assume the AI remembers things across sessions. It doesn't. They get frustrated, conclude the tool is limited, and either churn or adapt slowly through trial and error.

The pieces needed to fix this exist. They're just not assembled or surfaced.

What the Gap Looks Like in Practice A new user installs Claude Code, creates a CLAUDE.md, and starts working. By session 3 they're re-explaining their stack. By session 5 they're pasting the same context every time. The tool feels like a capable but amnesiac contractor. The user blames the model, not the missing infrastructure.

Proposed Solution

What's actually missing is a four-layer stack:

Persistent memory — typed, file-based memories (user, feedback, project, reference) that survive across sessions Session continuity — a SessionStart hook that reads a MEMORY.md index and injects relevant context automatically Cross-session recall — vector search over session history so the AI can answer "what did we decide about X?" without being told An inbox pattern — a drop zone (phone, notes app, anywhere) that feeds context into the next session asynchronously None of this requires new model capabilities. It's hooks, markdown files, and a local embedding model. The architecture is documented by Simon Willison and others. It works today.

Reference Implementation Over ~14 sessions building Portion (a Flutter + Riverpod iOS/macOS app), I assembled this stack incrementally:

~/.claude/memory/ — typed memory files with a MEMORY.md index, loaded on every session ~/.claude/hooks/session-start.sh — injects MEMORY.md content + INBOX items into session context memsearch plugin — local ONNX embeddings (bge-m3), no API key, searches .memsearch/memory/ across sessions Obsidian INBOX/ — iPhone drop zone; items ingested and archived at session start A wiki pattern (wiki/_index.md, wiki/projects/, wiki/_log.md) — LLM-maintained compiled knowledge base The result: every session opens with full context. No re-explaining the stack. Decisions persist. The AI behaves like a collaborator who was there last time, because effectively it was.

The Ask Add a guided setup flow — claude --setup or a first-run wizard — that bootstraps:

~/.claude/memory/ directory structure with starter templates A MEMORY.md index file A SessionStart hook that reads and injects MEMORY.md A project CLAUDE.md template with memory conventions documented Optionally: surface memsearch or an equivalent as a recommended plugin in the docs, with a one-command install path.

This is a 30-minute setup that most users will never discover on their own. Making it a guided 5-minute first-run experience would materially change the tool's value proposition — especially for long-running projects.

Why This Matters The limiting factor for AI-assisted development isn't the model — it's session amnesia. Fixing that at the infrastructure layer, not the prompt layer, is the right move. The pattern is proven. The reference implementation exists. The ask is modest: surface it, package it, make it the default starting point.

Alternative Solutions

No response

Priority

Medium - Would be very helpful

Feature Category

Interactive mode (TUI)

Use Case Example

No response

Additional Context

No response